tilelang.intrinsics.mma_macro_generator
=======================================

.. py:module:: tilelang.intrinsics.mma_macro_generator


Attributes
----------

.. autoapisummary::

   tilelang.intrinsics.mma_macro_generator.lift


Classes
-------

.. autoapisummary::

   tilelang.intrinsics.mma_macro_generator.TensorCoreIntrinEmitter
   tilelang.intrinsics.mma_macro_generator.TensorCoreIntrinEmitterWithLadderTransform
   tilelang.intrinsics.mma_macro_generator.INT4TensorCoreIntrinEmitter
   tilelang.intrinsics.mma_macro_generator.INT4TensorCoreIntrinEmitterWithLadderTransform


Module Contents
---------------

.. py:data:: lift

.. py:class:: TensorCoreIntrinEmitter(a_dtype = 'float16', b_dtype = 'float16', accum_dtype = 'float16', a_transposed = False, b_transposed = False, block_row_warps = 2, block_col_warps = 2, warp_row_tiles = 8, warp_col_tiles = 8, chunk = 16, reduce_k = 1, num_elems_per_byte = 1, is_m_first = False)

   Bases: :py:obj:`object`


   To eliminate Python syntax within TIR Macro.


   .. py:attribute:: M_DIM
      :value: 16



   .. py:attribute:: N_DIM
      :value: 16



   .. py:attribute:: WARP_SIZE
      :value: 32



   .. py:attribute:: dtype_abbrv


   .. py:attribute:: is_m_first
      :value: False



   .. py:attribute:: a_dtype
      :value: 'float16'



   .. py:attribute:: b_dtype
      :value: 'float16'



   .. py:attribute:: accum_dtype
      :value: 'float16'



   .. py:attribute:: a_transposed
      :value: False



   .. py:attribute:: b_transposed
      :value: False



   .. py:attribute:: block_row_warps
      :value: 2



   .. py:attribute:: block_col_warps
      :value: 2



   .. py:attribute:: warp_row_tiles
      :value: 8



   .. py:attribute:: warp_col_tiles
      :value: 8



   .. py:attribute:: chunk
      :value: 16



   .. py:attribute:: warp_rows
      :value: 0



   .. py:attribute:: warp_cols
      :value: 0



   .. py:attribute:: reduce_k
      :value: 1



   .. py:attribute:: threads
      :value: 128



   .. py:attribute:: num_elems_per_byte
      :value: 1



   .. py:method:: get_store_index_map(inverse = False)


   .. py:method:: extract_thread_binding(thread_id, is_m_first = None)

      is_m_first: True if the thread binding is in the form of (tx, warp_n, warp_m)
      which represents [warp_size, block_row_warps (split n), block_col_warps (split m)]
      Otherwise, it is in the form of [warp_size, block_col_warps (split m), block_row_warps (split n)]



   .. py:method:: ldmatrix_a(A_local_buf, A_shared_buf, ki, rk = 0)


   .. py:method:: ldmatrix_b(B_local_buf, B_shared_buf, ki, rk = 0)


   .. py:method:: mma(A_local_buf, B_local_buf, C_local_buf, k_inner = 0)


   .. py:method:: stmatrix(C_local_buf, C_buf, pid_m=None, pid_n=None)


   .. py:method:: make_mma_load_layout(local_buf, matrix = 'A')

      Create a layout function for storing MMA results into a fragment buffer.
      This layout is used in conjunction with `inverse_mma_store_layout` to
      map fragment indices to threads and local indices.

      :param local_buf: The local buffer representing a fragment of a matrix.
      :type local_buf: tir.Buffer

      :returns: A fragment object that describes how threads and indices
                in `local_buf` are laid out.
      :rtype: T.Fragment

      :raises AssertionError: If `local_buf` is not detected to be a fragment buffer.



   .. py:method:: make_mma_store_layout(local_buf)

      Create a layout function for storing MMA results into a fragment buffer.
      This layout is used in conjunction with `inverse_mma_store_layout` to
      map fragment indices to threads and local indices.

      :param local_buf: The local buffer representing a fragment of a matrix.
      :type local_buf: tir.Buffer

      :returns: A fragment object that describes how threads and indices
                in `local_buf` are laid out.
      :rtype: T.Fragment

      :raises AssertionError: If `local_buf` is not detected to be a fragment buffer.



.. py:class:: TensorCoreIntrinEmitterWithLadderTransform(a_dtype = 'float16', b_dtype = 'float16', accum_dtype = 'float16', a_transposed = False, b_transposed = False, block_row_warps = 2, block_col_warps = 2, warp_row_tiles = 8, warp_col_tiles = 8, chunk = 16, reduce_k = 1, num_elems_per_byte = 1, is_m_first = False, transform_kind_a = 0, transform_kind_b = 0)

   Bases: :py:obj:`TensorCoreIntrinEmitter`


   To eliminate Python syntax within TIR Macro.
   With Ladder Transform Plugin.


   .. py:method:: ldmatrix_a(A_local_buf, A_shared_buf, ki, rk=0)


   .. py:method:: ldmatrix_b(B_local_buf, B_shared_buf, ki, rk=0)


   .. py:method:: mma(A_local_buf, B_local_buf, C_local_buf)


.. py:class:: INT4TensorCoreIntrinEmitter(a_dtype = 'float16', b_dtype = 'float16', accum_dtype = 'float16', a_transposed = False, b_transposed = False, block_row_warps = 2, block_col_warps = 2, warp_row_tiles = 8, warp_col_tiles = 8, chunk = 16, reduce_k = 1, num_elems_per_byte = 1, is_m_first = False)

   Bases: :py:obj:`TensorCoreIntrinEmitter`


   To eliminate Python syntax within TIR Macro.


   .. py:method:: mma(A_local_buf, B_local_buf, C_local_buf)


.. py:class:: INT4TensorCoreIntrinEmitterWithLadderTransform(a_dtype = 'float16', b_dtype = 'float16', accum_dtype = 'float16', a_transposed = False, b_transposed = False, block_row_warps = 2, block_col_warps = 2, warp_row_tiles = 8, warp_col_tiles = 8, chunk = 16, reduce_k = 1, num_elems_per_byte = 1, is_m_first = False, transform_kind_a = 0, transform_kind_b = 0)

   Bases: :py:obj:`TensorCoreIntrinEmitterWithLadderTransform`


   To eliminate Python syntax within TIR Macro.
   With Ladder Transform Plugin.


   .. py:method:: mma(A_local_buf, B_local_buf, C_local_buf)